VesselVAE: Recursive Variational Autoencoders for 3D Blood Vessel Synthesis

TL;DR

VesselVAE tackles the challenge of producing realistic, diverse 3D blood vessel geometries with complex tree-like topologies. It introduces a recursive variational autoencoder (RvNN) that encodes and decodes binary vessel trees, learning a latent manifold for both topology and geometry; latent samples $z_s(x) ~ N(\mu, \sigma)$ can be decoded to synthesize new vessels. The model optimizes a beta-like objective $L = L_{recon} + \alpha L_{topo} + \gamma L_{KL}$, enabling joint reconstruction and topology prediction while shaping the latent distribution. Quantitative metrics show strong alignment with real data in radius ($0.97$), length ($0.95$), and tortuosity ($0.96$), and qualitative results demonstrate multi-branch vessels with smooth radii; limitations include occasional ultra-thin segments and mesh self-intersections. This framework advances 3D vascular geometry synthesis with potential impact on medical training, hemodynamic simulations, and forays into differentiable surface representations.

Abstract

We present a data-driven generative framework for synthesizing blood vessel 3D geometry. This is a challenging task due to the complexity of vascular systems, which are highly variating in shape, size, and structure. Existing model-based methods provide some degree of control and variation in the structures produced, but fail to capture the diversity of actual anatomical data. We developed VesselVAE, a recursive variational Neural Network that fully exploits the hierarchical organization of the vessel and learns a low-dimensional manifold encoding branch connectivity along with geometry features describing the target surface. After training, the VesselVAE latent space can be sampled to generate new vessel geometries. To the best of our knowledge, this work is the first to utilize this technique for synthesizing blood vessels. We achieve similarities of synthetic and real data for radius (.97), length (.95), and tortuosity (.96). By leveraging the power of deep neural networks, we generate 3D models of blood vessels that are both accurate and diverse, which is crucial for medical and surgical training, hemodynamic simulations, and many other purposes.

Figures (4)

• Figure 1: Top: Overview of the Recursive variational Neural Network for synthesizing blood vessel structures. The architecture follows an Encoder-Decoder framework which can handle the hierarchical tree representation of the vessels. VesselVAE learns to generate the topology and attributes for each node in the tree, which is then used to synthesize 3D meshes. Bottom: Layers of the Encoder and Decoder networks comprising branches of fully-connected layers followed by leaky ReLU activations. Notice that right/left Enc-MLPs of the Encoder only execute when the incoming tree requires it. Similarly, the Decoder only uses right/left Dec-MLPs when the Node Classifier predicts bifurcations.
• Figure 1: (a) Additional renders of blood vessels generated using VesselVAE. Our approach is capable of generating diverse and intricate vessel structures, including variations in thickness, branching patterns, and curvatures. These outputs closely resemble real anatomical structures and demonstrate the effectiveness of our neural network architecture and training procedures. (b) Limitations of our method: While VesselVAE is able to generate a wide variety of complex structures, it may occasionally struggle to reproduce realistic data. For example, the sample at the top of the figure features an extremely thin segment that may not occur in real blood vessels. Additionally, the mesh reconstruction algorithm employed by our method can sometimes produce vessels with self-intersections, which are not physically plausible in biological systems.
• Figure 2: Dataset and pre-processing overview: The raw meshes from the IntraA 3D collection undergo pre-processing using the VMTK toolkit. This step is crucial for extracting centerlines and cross-sections from the meshes, which are then used to construct their binary tree representations.
• Figure 3: (a) shows the histograms of total length, average radius and tortuosity per branch for both, real and synthetic samples. (b) shows a visual comparison among our method and two baselines wolterink2018bloodhamarneh2010vascusynth.